[Un article de The Conversation écrit par Mônica Macedo-Rouet – Professeure des universités en psychologie de l”éducation, CY Cergy Paris Université]
I have read the writings of master's students for over twenty years. This year, I noticed a massive increase in the number of work which included whole passages stylistically close to the texts produced by the generative AI. I spent many hours scrutinizing the reports of the Compilatio software (a tool designed at the start to fight against plagiarism, gradually adapted to AI), to verify the authenticity of bibliographic references, to do online research and sometimes even in printed works, in order to know if my students had written their texts themselves.
Indeed, at present, no tool makes it possible to determine with certainty if a text has been produced by the generative AI. Among the suspicious cases, I detected quotes to authors and bibliographic references not found on the net or in the university library. These occurrences known as “hallucinations” fully justified a request for explanations to my students. Their answers left me perplexed.
If the students mainly recognized having used AI, they did not see where the problem was. All sent me the articles they had “read” and “treated” in the context of their work. They justified the use of generative AI as a means of “reformulating [leurs] words “,” structure [leurs] ideas “,” improve syntax “,” illustrate the ideas of each author “,” save time rather than return to each article “, or” make the bibliography to [leur] place “. All of this seemed to them completely normal and acceptable.
More serious for me, whose profession is to educate information evaluation, when I asked them why the name of an author or the title of a review quoted in their text were different from those who appeared on the first page of the article they had transmitted to me, there was an upper shoulder.
Where did their perception come from that the quotation of sources was a detail in writing a writing on a research subject?
The role of sources in scientific writings … and in the texts generated by AI
The attitude of the students, made of a mixture of surprise (certainly possibly feigned) and frustration comes, in my opinion, from the upheaval brought by the generative AI to the status of information sources in the texts.
In a scientific text, the role of information sources is fundamental. The source corresponds to all the parameters that provide the reader on the origin of the information, such as the author, the date of publication, or the media. It gives indications on the institutional and disciplinary affiliation of an author, the editorial process prior to the publication of information, and other clues which make it possible to interpret the words and to judge its reliability.
However, if researchers are constantly based on these criteria to assess the credibility of a text, this is the object of a learning process for students. In a precursor article on the subject, Wineburg compared the reasoning of historians and students of terminal on a set of documents about a controversial historical event. The source was the first criterion used by historians to assess the relevance and reliability of a document, while it only appeared in third position for high school students, who focused more on the content and readability of the texts. These results have been replicated in many studies.
Recently, everyone has been able to measure their importance in the context of the dissemination of false information on COVID-19. Without the source, the credibility of scientific information can hardly be assessed.
AI can contradict its sources: it is not designed to be faithful
In the texts generated by AI, the role of sources is significantly different.
Based on this technology, there is a corpus of gigantic sources which allows statistical models of language to learn and generate coherent texts and probably similar to the texts produced by humans.
But the sources only serve asinput During training and are not used as an explicit criterion of reliability when generating an answer. The model predicts the most likely text of a text, word by word, according to the regularities learned, without assessing the veracity of the information in relation to authenticated documents.
Thus, one can end up with a text generated by the perfectly coherent and nevertheless erroneous AI. Even when Chatgpt is asked to summarize a scientific article, it is necessary to verify that the information corresponds to that of the original article. Without a scrupulous verification of the texts produced by AI, there is a risk of reproduction of imprecise or incorrect information, and of attribution of certain ideas to false authors, which constitutes fraud punishable by sanctions.
Not citing your sources (correctly) is liable to sanctions
Students do not necessarily have the impression of cheating when they use AI as a writing help, because the texts generated by AI are not a plagiarism in the literal sense. In France, the Ministry of Higher Education and Research spoke on this subject in a answer to the question of a senator in 2023 (page 5289): “The works created by AI are not protected in themselves unless they reproduce works of the mind within the meaning of the Intellectual Property Code[…][donc] Copying a text produced by Chatgpt cannot be penalized with regard to the provisions of articles L. 122-4 and L. 335-2 of the Intellectual Property Code. »»
However, the same precise answer as: “The indication of sources is a legal, academic and ethical obligation. From an academic point of view, in particular, it must make it possible to assess the educational value of the original work carried out by its author. Do not mention the sources to make their work carried out by others or by an AI is, consequently, constituting fraud likely to be prosecuted and sanctioned, for users of higher education, in application of the provisions of articles R. 811-1 and following of the education code. »»In other words, using a text generated by AI does not exempt the student from properly quoting their sources. The sanctions can go as far as the university and the withdrawal of the diploma, without limitation period.
In short, not quoting your sources in the rules of the art is a fault that can have serious consequences for the schooling of a student, not to mention the fact that the simple copy of a text produced by AI does not guarantee learning. Because it requires active information processing on the part of the learner.
Everyone must therefore ensure that the sources used in their work are properly cited, according to the bibliographic and scientific standards in force. Unfortunately, these standards are sometimes taught too briefly or superficially – when they are – in the methodology of research at university.
A first track: to improve the detection of texts produced by AI
A first track would be to improve the detection of texts produced by AI.
Automatic detection software becomes more and more efficient in this task, but generative AI models are also improving in the application of paraphrase and “humanization” strategies of texts, which make automatic detection more difficult. In addition, some researchers strive to build models aimed at directly preventing automatic detection of the texts generated by the AI.
It is therefore an extreme and unequal balance of forces that is playing and may reproduce permanently, making it difficult to provide teachers of high -performance automatic detection tools.
To improve the detection of the texts generated by AI, a study not yet published, filed in the Arxiv platform, proposes to call on experts in the use of AI. Researchers have indeed observed that these experts are able to use several evaluation criteria in a flexible manner: “Vocabulary of AI”, presence of syntactic and stereotypical structures, absence of orthographic and grammatical faults, among others. These results obviously require being confirmed by a publication and replicated, but they suggest that it may be useful to train teachers in the application of these criteria.
Form – Always form! – the evaluation of information sources
Beyond the purely “detection” aspect of the texts, it is knowledge on the structure and the rhetoric of the texts generated by the AI that should be explained in order to integrate them into university pedagogy.
AI can help teachers and students in many tasks, but it cannot completely replace human judgment. The ethical use of AI is not limited to prohibiting certain processes or promoting the technical skills of students and teachers (for example, “how to make a good prompt?”). It goes beyond normative and technical aspects, and includes questions of epistemology, documentary knowledge, and metacognition essential to any controlled approach to finding information.
I am sure that we would benefit from having more open discussions with students about AI uses in university work, what it means for them and for us, teachers and researchers, and how we could take advantage of this incredibly powerful technology without renouncing our ethical principles or our intellectual training ambitions. It would be a debate in the name of knowledge, learning and truth, a debate that our university and our democracy need so much.
With an unwavering passion for local news, Christopher leads our editorial team with integrity and dedication. With over 20 years’ experience, he is the backbone of Wouldsayso, ensuring that we stay true to our mission to inform.
